CMOS Image Sensors - Past Present and Future Boyd Fowler, Xinqiao(Chiao) Liu, and Paul Vu Fairchild Imaging, 1801 McCarthy Blvd. Milpitas, CA 95035 USA Abstract sors. CMOS image sensors can integrate sensing and process- In this paper we present an historical perspective of CMOS ing on the same chip and have higher radiation tolerance than image sensors from their inception in the mid 1960s through CCDs. In addition CMOS sensors could be produced by a vari- their resurgence in the 1980s and 90s to their dominance in the ety of different foundries. This opened the creative flood gates 21st century. We focus on the evolution of key performance pa- and allowed people all over the world to experiment with CMOS rameters such as temporal read noise, fixed pattern noise, dark image sensors. current, quantum efficiency, dynamic range, and sensor format, The 1990’s saw the rapid development of CMOS image i.e the number of pixels. We discuss how these properties were sensors by universities and small companies. By the end of improved during the past 30 plus years. We also offer our per- the 1990’s image quality had been significantly improved, but spective on how performance will be improved by CMOS tech- it was still not as good as CCDs. During the first few years of nology scaling and the cell phone camera market. the 21st century it became clear that CMOS image sensor could out–perform CCDs in the high speed imaging market, but that Introduction their performance still lagged in other markets. Then, the devel- MOS image sensors are not a new development, and they opment of the cell phone camera market provided the necessary are also not a typical disruptive technology. MOS image sensors capital to improve CMOS image sensors to a point where they were first devised in the mid 1960’s, when the incumbent tech- out–perform CCDs in many applications today. nology for electronically capturing video was the vidicon tube Another good description of the history of MOS/CMOS not the CCD. Vidicon technology was mature with more than image sensors is presented by Fossum in [2, 3]. 30 years of television development behind it. Although the im- In the remainder of this paper we will elaborate on the age quality generated by a vidicon tube was excellent, its size, history of MOS/CMOS image sensors using key performance weight, and lag were disadvantageous for many applications1. parameters to track their development. In the next section we These limitations created the impetus necessary to develop solid present definitions for temporal read noise, fixed pattern noise, state MOS image sensors. dark current, quantum efficiency, dynamic range, and format as During the 1960’s MOS image sensors were used in many they pertain to MOS image sensors. In following three sections applications including optical character recognition, reading we chronicle the invention, resurgence and the present and fu- aids for the blind, and low resolution 2D cameras. Unfortu- ture of MOS image sensors. Finally we present our conclusions. nately MOS sensors had many problems including both high fixed pattern and temporal noise. CCDs were invented in 1969 Definitions and quickly offered better imaging performance than MOS im- In this Section we discuss six key parameters that deter- age sensors. This forced MOS image sensors into only a few mine an image sensor’s performance. This is clearly an over- specialized applications such as spectroscopy. simplification, because more than six parameters are required to Throughout the 1970’s CCDs dominated the image sen- completely characterize an image sensor. In this paper the six sor market, due to their inherently lower FPN. During the early parameters that we will focus on are temporal read noise, fixed 1980’s CCDs were controlled by only a few companies and were pattern noise, dark current, quantum efficiency, dynamic range, produced for only a few applications. This made conditions and sensor format. right for universities and companies to start the exploration of In order to simplify the explanation of these parameters we CMOS image sensors, the next generation of MOS image sen- present a simplified model for an image sensor, i.e. sors, for various applications. This included applications such as ( )= (η + dc + ( )) + , Yi, j t gi, j Fi, jtint Ii, j tint Ni, j t Oi, j (1) machine vision, hand held video cameras,and space based sen- where Yi, j is the sensor output voltage from pixel (i, j), η is the 1 For a detailed description of vidicon tubes see [1]. quantum efficiency of the sensor, Fi, j is the photon flux at pixel , dc (i j), tint is the sensor integration time, Ii, j is the sensor dark architecture as shown in Figure 1 that used a reverse biased pho- current at pixel (i, j), Ni, j is the temporal read noise, gi, j is the todiode to integrate photogenerated charge, connected to a sin- conversion gain of pixel (i, j), and finally Oi, j is the fixed offset gle pchannel MOS transistor to individually access each pixel in voltage of pixel (i, j). a linear array. The photodiode is read out by turning on the MOS We define temporal read noise of pixel (i, j)astheRMS access transistor and transferring charge from the photodiode to variation in Ni, j(t). To measure this value the sensor integra- a column amplifier via the column bit line. During this process tion time is typically set to zero, i.e. tint = 0andYi, j(t) is mea- the photodiode is also reset to fixed voltage determined by the sured multiple times. Temporal read noise is typically reported column level amplifier. Typically a capacitive transimpedance in electrons(e-) RMS. Temporal read noise consists of thermal amplifier is used at the column level for passive pixel sensors. and 1/f noise from transistors in the readout path, and detector Between readout periods the photodiode is always integrating (photogate or photodiode) reset noise[4]. photocharge. This development dramatically increased the sen- We define fixed pattern noise (FPN) as the pixel to pixel sitivity of MOS image sensors compared with non-integrating variation of the sensor when the input signal Fi, j is constant for photoresistive sensors. all pixels. In our simple model this implies that FPN consists of Quickly, researchers at Plessey, Stanford University, Uni- two components, offset and gain. The offset FPN is caused by versity of Waterloo, and Nippon Electric Company started de- dark current variation and transistor mismatch, i.e. the pixel to velopment of MOS image sensors. In 1968 P. Noble at Plessey ( dc )+ pixel variations in gi, j Ii, j tint Oi, j. Gain variation is caused in Northamptonshire England proposed the MOS active pixel by variation in effective pixel size and conversion gain, i.e. gi, j. architecture[8]. This pixel architecture, shown in Figure 2, uses We will define offset FPN as the peak to peak variation from a pchannel MOS transistor to buffer the photodiode voltage. pixel to pixel divided by the sensor full scale output, and we This circuit trades–off higher sensor speed and lower temporal define gain FPN as maxgi, j − mingi, j /gi, j where gi, j is the read noise for smaller pixel fill factor, i.e. the ratio of photodi- average sensor gain. See [5] for detailed discussion of FPN in ode area to the pixel area. The photodiode is reset by turning MOS image sensors. on M1, which forces the voltage across the photodiode to VDD. dc Dark current Ii, j is the photodetector leakage current. It The photodiode voltage is readout by turning on M3 and using is typically measured as a function of integration time and M2 as a source follower. Just like the MOS passive pixel archi- temperature when Fi, j = 0. It is often reported in units of tecture, between readout/reset periods the photodiode is always pA/cm2 at a specific operating temperature. Dark current con- integrating photocharge. Other researchers at Plessey and at the sists of both Shockley–Hall–Read (SHR) generation and diffu- University of Waterloo in Canada also worked on the MOS ac- sion currents[6]. It is very sensitive to temperature and silicon tive pixel architecture[9, 10, 11] during late 1960’s and early defect density. For example, the dark current doubling rate of a 1970’s. MOS image sensor is typically between 6-12◦C at room temper- At the same time researchers at Stanford University were ature. working on MOS passive pixel image sensors for aiding the Quantum efficiency η is defined as the ratio between the blind to read standard printed text[12, 13, 14]. This research number of collected electron/hole pairs to the number of inci- was commercialized into the Optacon in the early 1970’s. dent photons at a specific wavelength. The quoted quantum effi- In Japan, Nippon Electric Company (NEC) was also ciency is typically the maximum achieved over the visible spec- developing MOS passive pixel image sensors for line scan trum. applications[15] in the early 1970’s. Dynamic range is defined as the maximum non-saturating The performance of early MOS image sensors was limited signal divided by the temporal read noise. For example, if the by offset fixed pattern noise caused by MOS transistor threshold average sensor conversion gain gi, j = 2µV/e- and the maximum and gate–to–source overlap capacitance variation. For example output signal is 1V and the sensor temporal read noise is 50e- a typical MOS image sensor had about 5% full scale (FS) offset RMS, then the dynamic range of the sensor is 500000/50 = fixed pattern noise.